* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download A semi-automatic resolution of anaphora and ellipsis in a large
Malay grammar wikipedia , lookup
Macedonian grammar wikipedia , lookup
Antisymmetry wikipedia , lookup
Navajo grammar wikipedia , lookup
Udmurt grammar wikipedia , lookup
Sanskrit grammar wikipedia , lookup
Zulu grammar wikipedia , lookup
Chinese grammar wikipedia , lookup
Lithuanian grammar wikipedia , lookup
Esperanto grammar wikipedia , lookup
Swedish grammar wikipedia , lookup
Modern Greek grammar wikipedia , lookup
Arabic grammar wikipedia , lookup
Old English grammar wikipedia , lookup
Japanese grammar wikipedia , lookup
Georgian grammar wikipedia , lookup
Portuguese grammar wikipedia , lookup
Kannada grammar wikipedia , lookup
Modern Hebrew grammar wikipedia , lookup
French grammar wikipedia , lookup
Yiddish grammar wikipedia , lookup
Spanish pronouns wikipedia , lookup
Latin syntax wikipedia , lookup
Turkish grammar wikipedia , lookup
Scottish Gaelic grammar wikipedia , lookup
Ancient Greek grammar wikipedia , lookup
Romanian nouns wikipedia , lookup
Romanian grammar wikipedia , lookup
Sloppy identity wikipedia , lookup
Serbo-Croatian grammar wikipedia , lookup
Lexical semantics wikipedia , lookup
Spanish grammar wikipedia , lookup
Pipil grammar wikipedia , lookup
Submitted to the special issue of Computational Linguistics (on resolution of anaphora and ellipsis) Eva Hajičová, Jarmila Panevová and Petr Sgall A semi-automatic resolution of anaphora and ellipsis in a large corpus of Czech Faculty of Mathematics and Physics, Charles University Malostranské nám. 25, 11800 Prague 1, Czechia {hajicova,panevova,sgall}@ufal.mff.cuni.cz ABSTRACT The Prague Dependency Treebank represents a semi-automatic three-layer annotation system, in which the layers of morphemic and 'analytic' (surface-syntactic) tagging are followed by the layer of tectogrammatical tree structures, which represent the meaning of every sentence in its environment, taking into account its topic-focus articulation. The inflectional complexity of Czech has led to a strategy with which certain types of coreference (including ellipsis) are accounted for 'manually', and others by a subsequent automatic module that specifies the reference of reflexive pronouns and restores deleted nodes in coordinated constructions. The handling of textual coreference is made possible by specific attributes which reflect the linking of sentences to each other. The changes of the degrees of activation will be registered as far as derivable from the use of nouns in a discourse. 1. Introductory remarks A specific strategy is necessary for the resolution of anaphora and ellipsis, so that a theoretical background based on a comparison of typologically different languages is highly useful. In the present paper we concentrate on the following aspects of handling coreference:1 (i) In the annotation scheme of the prague dependency tree-bank our paper is based on not only a surface-syntactic tagger is included, but rather a procedure leading from the outer form of written text to underlying representations, which should be suitable as an input for a semantico-pragmatic interpretation in the sense of intensional logic. The handling of spoken discourse is connected with problems of its own, but can well make use of many parts and aspects of the present procedure. (ii) The main factors useful for anaphora and ellipsis resolution are identified and discussed, connected both with the coherence of a discourse (degrees of salience and their changes during communication) and with grammatical conditions (reflexive and relative items, control, etc.). Moreover, a way to achieve a fully automatic annotation, incorporating both structural and statistical methods, is indicated as an outlook. (iii) The underlying linguistic theory is based on dependency syntax and on a fundamental distinction between syntax and morphemics (both synthetic and analytic; see Sgall et al. 1986, Sgall 1992, Hajičová 1993, Petkevič 1995). This combination allows for a maximally economical description of different layers of sentence structure, including the topic-focus articulation (TFA), which is neglected in many theoretical descriptions, although it is relevant not only for the contextual potential of the sentence, but also for its semantics. As will be shown, TFA belongs to the factors relevant for the resolution of contextual anaphora and ellipsis. (iv) The strategy elaborated on the basis of this theory, based both on grammatical and on textual factors, may be used with advantage not only for Czech and other inflectional (fusional) languages, but also as a general basis of formal and computer-oriented linguistic description, appropriate for languages of different types and for their comparison. Thus it may be of interest to compare the possibilities offered by the present descriptive framework with those proper to the majoritary approaches based on constituency. (v) The corpus with annotations including data on anaphora and ellipsis resolution will allow for a direct and complex checking of the descriptive framework used and of the linguistic theory on which it is based. If the results then show that such an economical descriptive framework is suitable for the objectives enumerated in the previous points (and for other aims), the importance of this challenge will become indisputable. Our paper has the following structure: in Section 2 we attempt to specify the place of coreference resolution in the complex tagging procedure. The strategy of handling coreference is presented in Section 3.1; a distinction is made between methods used for coreference determined by grammatical and by textual factors (Section 3.2 and 3.3, respectively), and for handling certain specific aspects of ellipsis (Section 3.3). Some prospects for future work in this domain are sketched in Section 4. 2. Coreference in the semi-automatic tagging of Czech 2.1. The strategy of semi-automatic tagging In the domain of coreference and ellipsis, the typological differences between English and a language with rich inflection, such as Czech, are clearly to be seen. The rich morphemics allows for a coreferential item to be deleted in many cases in which an ellipsis is impossible in the English text. To put it in an extreme way, in principle everything in any position can be deleted in Czech if it is identifiable on the basis of the context; this concerns not only the prodrop character of Czech, but often the deletion of the whole topic of a sentence (or those parts of the topic that are not contrastive). In English the situation differs, since the grammatical restrictions on ellipsis are stronger, requiring the presence of auxiliaries and pronominal or similar items, cf. the examples (1) - (3): (1) (Martin byl včera ve škole.) Marie taky. lit.: '(Martin was yesterday at school.) Mary too.' [Mary was there too.] (2) (Martin včera dobře spal.) Marie taky. lit.: '(Martin yesterday well slept.) Mary too.' [Martin slept well yesterday.] So did Mary. (3) (Milena to řekla Pavlovi,) Pavel Karlovi a ten Petrovi. lit.: '(Milena told it to-Paul,) Paul to-Charles, and he to-Peter.' In Czech, ellipsis thus clearly represents a specific kind of coreference, and this may be assumed to hold also for many other languages: it may be supposed that the speaker deletes only those items the referents of which are salient enough at the given time-point of the discourse to be easily accessible for the hearer; cf. the end of Section 2.2 below. Up to now the resolution of ellipsis and anaphora can only be handled semi-automatically in the procedure of the syntactic tagging of the Czech National Corpus, which is being carried out at Charles University, Prague. The large annotated corpus that is being built in this way (the Prague Dependency Treebank, PDT) can be used as a set of training data for further automatic treatment, as soon as the range of the data is extensive enough to allow for a successful application of statistical (or hybrid, for that matter) methods. The set of empirical data that are available already now is rather large, since the Czech National Corpus (CNC), compiled and processed in the Institute of CNC (led by F. Čermák) at the Faculty of Philosophy, Charles University, Prague, now comprises more than 100 million word tokens from different kinds of texts. A part of CNC has been used as the basis of the Prague Dependency Treebank (PDT, see Hajič 1998), conceived of as a semi-automatic three-layer annotation system (see Hajičová 1998), in which the layers of morphemic and 'analytic' (surface-syntactic) tagging are followed by a third layer, viz. that of tectogrammatical tree structures (TGTSs in the sequel); TGTSs differ from theoretically substantiated tectogrammatical representations (which are multidimensional, due to the free combinability of dependency and coordination relations) in that they have the form of bidimensional trees, with coordinating conjunctions handled as specific nodes. An automatic procedure of morphemic tagging has been formulated (see Hajič and Hladká 1997; Hladká 2000). Moreover, by now, 100 000 sentences from the Corpus have obtained their 'analytic' annotations, i.e. tree structures in which every occurrence of a word form and of a punctuation mark obtains a corresponding node with a surface syntactic label. We expect to get several thousands of sentences annotated by their TGTSs before the end of the year 2000. The TGTSs are intended to represent the underlying syntactic structure of sentences, which would be appropriate as the input to semantic(-pragmatic) interpretation, since the irregularities of the shallow layers, including synonymy and ambiguity, are absent on this level (see Hajičová 1993 and the writings quoted there). This implies that in the TGTSs, nodes for cases of (surface) deletions should be added ('reconstructed'). Hundreds of sentences have already been tectogrammatically tagged as for the main points, including the restoration of most of the deleted items. A more detailed annotation has been achieved, up to now, for about 100 sentences (see below on the difference between the 'large corpus', LC, and the 'model corpus', MC). A theoretically substantiated labelling of the TGTSs can be gained in this way, distinguishing between different kinds of objects and of adverbials, between meanings of function morphemes, topic and focus, and so on. The result will be much more complex than that of a parser or tagger of the usual kinds: not only the grammatical well-formedness will be checked, but disambiguated representations of sentences will be achieved, which (although underspecified in the points in which the sentence structure is not fully specific indistinctness, "systematic ambiguity", scopes of quantifiers) would constitute an appropriate input for a procedure of semantico-pragmatic interpretation. 2.2. Automatic and intellectual steps The units of annotation in PDT are sentences as occurring in the texts in the CNC, and the human annotators are instructed to assign every sentence a (disambiguated) structural description according to the meaning of the sentence in its environment. Several parts of the tagging procedure have been (or are being) formulated as general steps, carried out automatically. Two of them concern the transduction from morphemic annotations and 'analytic' (surface syntactic) trees (ATSs) to the underlying ones (tectogrammatical tree structures, TGTSs): (i) In an automatic 'pre-processing' module, the ATSs are pruned, i.e. the nodes that are marked as function words in the ATSs get deleted. The values of Tense, Aspect, Gender, Number and degrees of comparison carried by the function words or by morphemic tags corresponding to inflectional endings and affixes get the position of indices within the labels of lexical nodes, i.e. of values of 'grammatemes' (some asymmetries between forms and their respective functions will be solved later during the manual procedure). The grammateme of Sentential Modality (with the values ENUNC, INTERR, IMPER, DESID) is specified automatically with all heads of main clauses on the basis of the node standing for the final sentence boundary and of other data (esp. particles) present in the analytic tree. The analytic function Subject with an active verb is converted into the tectogrammatical functor ACT (actor/bearer); whenever possible, every analytic word form is combined into a single node, the label of which contains the lexical value (lemma) indexed with a string of grammatemes derived from endings and grammatical affixes, as well as from auxiliary verbs, articles, prepositions, conjunctions; the only exceptions are coordinating conjunctions, which retain their nodes as governors of the coordinated syntagm. (ii) The 'manual' part of underlying tagging consists of operations that can only be performed intellectually, before further analysis helps to find reliable criteria to identify specific contexts in which secondary functions occur. Along with specifying most of the syntactic relations (functors) and the more difficult cases of grammatemes (including those reflecting the topic-focus articulation of the sentence and corresponding movements), the following operations important for the phenomena of coreference are included here: (a) Nodes for the deleted items are restored and their attribute 'lemma' (to be filled by the lexical value) either gets the value of a pronoun (including specific symbols for a 'General Participant' and for a 'Controllee' with control verbs), or the attribute 'lemma' is left vacant for further treatment (e.g. in coordinations, see point (i) above. (b) With possessive adjectives and pronouns dependent on nouns, the number and gender values of the bases they are derived from are of an anaphoric nature, due to their agreement with the values of their antecedent) are taken as the values of their respective grammatemes. Note that these values are of an anaphoric nature, due to their agreement with the values of their antecedent. The following situations may serve as an illustration: jeho 'his, its' gets the values SING, ANIMATE (or INANIMATE or NEUTER, according to the context, i.e. to the gender of the antecedent), její 'her' gets SING, FEM, jejich 'their' gets PLUR and the appropriate gender, můj 'my' gets SING and ANIM, FEM, or NEUT, matčin 'mother's' gets SING, FEM, and so on. The help of a 'user-friendly' software enables the annotators to work with diagrammatic shapes of trees. (iii) After the 'manual' handling of TGTSs, another automatic module adds information that can be 'retrieved' automatically at that stage; among other tasks, this module makes further additions concerning coreference, as for instance: (a) the lemma of the node carrying the functor value ACT is assigned to the attribute COREF of an occurrence of the reflexive pronoun se that has not yet been treated (i.e. the PAT Patient, Objective - of an active verb); (b) the remaining nodes without lemmas (in coordinated constructions or in apposition) are assigned the lemmas of their counterparts in the given construction; e.g. in (4) the node corresponding to the deleted second occurrence of the verb (which has been added intellectually as governing both Karel.ACT and Milenu.PAT) gets a lemma identical to that of the lefthand coordinated item pozval 'invited'. (4) Jirka pozval Marii a Karel Milenu. lit. Jirka invited Mary and Karel Milena. In the outlook, the automatic procedure is supposed to be enriched in various respects, to cover at least the most regular phenomena of several further subdomains, among which it is directly relevant for textual coreference that the development of the degrees of salience (activation) of the 'stock of shared knowledge' (see Hajičová 1993; Hajičová et al. 1981; 1995) will be registered as far as derivable from the use of nouns in subsequent utterances in a discourse. It will then be possible to specify automatically the entity much more salient than every other item and thus available to be referred to by a weak pronoun (or even by an ellipsis). This specification will then make it possible to determine the value of the attribute COREF in case of a pronominal or even 'elliptical' reference. 2.3. Large corpus and model corpus The annotation on the underlying syntactic layer (resulting in the TGTSs) is carried out in parallel in two streams both having as their inputs the result of the automatic preprocessing of the analytic syntactic trees (for a description of this procedure, see Böhmová and Hajičová 1999). The outputs of these streams differ in the size of data and the size of information carried by the tags: (i) The set of "core" TGTSs (called 'large corpus', LC) has a large size, and it is being annotated with a higher speed and with tags carrying information about (a) the types of dependency relations and (b) values indicating the topic/focus articulation. (ii) The set of "full" TGTSs (the 'model' corpus, MC) has a smaller size, being annotated with a lower speed and with tags carrying complete tectogrammatical information (for a detailed characteristics of TGTSs, see Hajičová et al. 1999). 3. Coreference 3.1. Strategy of coreference resolution Since the PDT is conceived of i.a. as a resource for linguistic research beyond the limits of the sentence, three specific contextual attributes have been introduced in the TGTSs reflecting the linking of sentences to each other and to the situational context: (i) the attribute COREF having as its value the lexical value of the antecedent of the given anaphoric node (this node itself may be present on the surface, or deleted); (ii) the attribute CORNUM with a value equal to the serial number of the antecedent of the given node (which is needed in case of the same noun or pronoun occurring twice in the sentence), and (iii) the attribute CORSNT indicating whether the antecedent is in the same sentence (the value NIL) or in the preceding context (the value PREV). In the case of grammatical coreference (see below, Sect. 3.2), the substantial feature of which is the presence of the antecedent in a specified syntactic position of the sentence, an additional attribute ANTEC is used with the value equal to the dependency relation (functor) of the antecedent of the given node. The choice of the lexical values of the restored nodes is governed by the following principles: If an anaphoric node deleted on the surface is being restored, its lexical value is specified as an anaphoric (weak) pronoun (P in the sequel), a specific lexical value (L), or a technical value (such as Cor for the 'controllee'). L is assigned only to such specific restored nodes as those in coordinated constructions (see the discussion of ex. (37) in Section 3.4 below). Otherwise, wherever the lexical unit of the antecedent can be specified in a non-systemic way, yet univocally (it may be carried over from the preceding sentences, crossing the full stop), the lemma of the antecedent is added to the label of the relevant node only as the value of the grammateme COREF, in the MC; thus with ex. (5), if it is clear that he did not give a bunch of flowers e.g. to Mary, whose image has been made highly salient by the context, the pronominal form jí 'her' (a form of the pronoun on 'he') will be assigned the label 'on.FEM.SG.ADDR', the value of the attribute COREF of this node will be Jiřina and that of CORNUM: 3. (5) Přišel k Jiřině a dal jí kytku. 'He came to J. and gave her a flower.' With deleted verbs full lexical values must be supplied, since Czech does not have a pronominal verb such as the English do. If the antecedent of a given node is pronominal, then, of course, a pronoun is added as the lexical label of the restored node. The system of annotation of the TGTSs makes it possible to reflect the distinction between grammatical and textual coreference (see Panevová 1991). A typical example of the former is the coreference of the subject of the infinitival complementation of the control verbs (the subject of the infinitive gets the lexical value Cor) and the coreference of the reflexive pronouns (with which the value of COREF is identical to the lemma of the subject), as well as that of the relative words in their relationship to their antecedents. With textual coreference (e.g. the 'deleted' pronominal subjects in Czech as a pro-drop language or other cases of pronominal reference) the nodes for the anaphoric expressions get P both as their lexical value and as the value of COREF. Although also nouns, verbs, etc., can have a coreferential value, which we plan to reflect in the future form of the procedure, we do not discuss these cases in the present paper. 3.2. Grammatical coreference With grammatical coreference, the value of COREF is filled in (by the lemma of the controller or of another antecedent); the lemma and the functor of the coreferring node are determined both in the LC and in the MC. In MC, also the values CORNUM and ANTEC are specified (with CORSTN, the unmarked value NIL remains, since with grammatical coreference the antecedent always occurs in the same sentence). The typical cases of grammatical coreference are reflexive and relative pronouns, and control: (i) Reflexives: The forms se, si, sebe, sobě, sebou (case and gender forms of 'himself') and svůj, svá, etc. (the possessive reflexive), have a common lemma se. With active clauses, the value of COREF (i.e., the lemma of the subject), will be added in the second phase of the automatic procedure, both in the MC and in the LC; with svůj and with passive both the lemma and COREF are supplied manually. The reflexive pronoun (or particle) has different functions; most often the following values occur: se - PAT, ACT; in certain cases just a lexical derivative is present, e.g. in smát_se 'laugh', šířit_se 'spread'; si - ADDR (cf. (6)), PAT, BEN(efactive), lexemes such as pospíšit_si 'hurry up', or ETHD ('ethical dative'), e.g. in dělejte si, co chcete 'do whatever you wish'; svůj - se.APP, with Gender and Number of its antecedent (6) Koupil si.ADDR jízdenku do Vídně. 'He bought for-himself a ticket to Vienna.' A specific case is that of the reciprocal use of se, si, sebou...; in the LC reciprocity is disregarded, but in the MC the pronoun gets the lemma se-Recp; most often the relation of reciprocity is constructed as coordination, and then it is the lemma of the conjunction that appears in COREF: (7) Honza a Jirka se střídali. 'Johnny and George were alternating with each other.' Honza.ACT a.CONJ Jirka.ACT se_Recp.PAT;COREF:a střídali. Clauses with a plural subject are handled in a similar way: (8) Chlapci se střídali. 'The boys were alternating with each other.' Chlapci.ACT se_Recp.PAT;COREF:chlapec střídali. In the MC, the attributes CORNUM and ANTEC get the number and the functor, respectively, of the antecedent. (ii) Relative clauses are handled as congruent adjuncts of their antecedents; the functor of their verbs mostly is RSTR or DES (for restrictive and non-restrictive or 'descriptive' adjunct, respectively), both in LC and MC; the relative word gets its functor in accordance with its syntactic role within the clause, and the values of its attributes COREF, CORNUM in the MC correspond to the lemma and the number of the antecedent, respectively: (9) Jsou to lidé, kteří mají podobné názory. They are people who have similar opinions Jsou to lidé, kteří.ACT;COREF:lidé;CORNUM:3 mají.RSTR podobné názory. (10) Jsme lidé, kteří se liší od zvířat... We are people, who differ from animals...' Jsme lidé, kteří.ACT;COREF:lidé;CORNUM:2 se_liší.DES od zvířat... Relative adverbs can have different functions, e.g. that of a Directional, not necessarily identical with that of the clause as a whole or with the anaphoric word accompanying it (and treated as its head), cf. (11), with the values specified in (11'): (11) Kam to dáš, tam to najdeš. where it you-put there it you-find 'Where you put it there you find it.' (11') Kam.DIR-where_to;COREF:tam;CORNUM:4 to dáš.RSTR, tam.LOC;COREF:dáš;CORNUM:3 to najdeš. (iii) The relation of control is handled manually up to now, although a part of the task is supposed soon to be fulfilled automatically. With most verbs of control the controller is specified as their Actor, Addressee or Patient. Due to the intrinsically syntactic character of the function of controller, we prefer to resolve its relationship to the obligatorily deleted controllee in the form of restoring the latter as a node equipped with the 'technical' lemma Cor; in LC it just gets the functor value ACT (with a passive infinitive, this value is either PAT or ADDR, according to the valency data of the controlling verb); the lemma of the controller is indicated as the value of COREF in the label of the restored node of the controllee. In the MC also the functor of the controller and its position are indicated in the label of the controllee (as the values of ANTEC and CORNUM, respectively); see the following examples (with additions within LC placed in brackets; the MC forms of (12) and (13) are marked by primes: (12) Podnik plánoval [Cor.ACT.ELID;COREF:podnik] zvýšit výrobu. The firm planned to step up (its) production. (12') Podnik plánoval [Cor.ACT.ELID;COREF:podnik;CORNUM:1; ANTEC:ACT] zvýšit výrobu. (13) Radili synovi [Cor.ACT.ELID] k odchodu. lit.: They advised their son for departure. (13') Radili synovi [Cor.ACT.ELID;COREF:syn;CORNUM:2; ANTEC:ADDR] k odchodu. Note: We distinguish: Jirka slíbil přijít George promised to come , where a node with the lemma Cor functions as ACT of the infinitive (since the alternative that someone else would be coming is out of question) from a structure with textual coreference in: Jirka slíbil, že přijde G. promised that he would come , where as ACT of the infinitive the personal pronoun on with gender assigned according to context is supplied, because on 'he' is 'free' in its interpretation (it may but need not refer to Jirka. As (13) shows, also nouns of action functioning as objects of a verb of control are treated in this way. This concerns also the so-called Slavonic infinitive with accusative (the verb slyšet 'hear' has the frame ACT PAT (EFF) on this reading, i.e. the position of the 'second object', of Effect, occupied by an infinitive, is optional, although, if present, obligatorily connected with control): (14) Honza slyšel Karla [Cor.ACT.ELID;COREF:Karel;] otvírat dveře. Johnny heard Charles open the door (14') Honza slyšel Karla [Cor.ACT.ELID CORNUM:3;ANTEC:PAT] otvírat dveře. If the position of PAT with the governing verb is not occupied by a specific lexeme in this construction, then it gets the lexical value of Gen (a general participant), cf. (15): (15) Slyšel [Cor.ACT.ELID;COREF:Gen] otvírat dveře. He heard the door open. 3.3. Textual coreference The textually coreferring node, which either corresponds to a pronoun or is a case of restored deletion, obtains a functor and a P lemma both in the MC and in the LC. Also coreferentially used nouns are handled in this way, at least in those cases in which their anaphoric function is explicitly expressed by an accompanying pronoun, as e.g. in ta kniha 'that book'. In the MC, the grammateme values of the antecedent are added to the textually coreferring node: its attribute COREF obtains as its value the lemma of the antecedent, CORNUM gets the value of the serial number (according to its word-order position, adjusted by decimal fractions in case of preceding deletion restorations); in CORSTN the unmarked value NIL is placed automatically, and changed into PREV if the antecedent is in the preceding sentence. In the LC, the attribute COREF is not filled in, and if the relevant node has been deleted, it is restored only in the case of a zero subject or of another deleted obligatory participant; if the head of a deleted item has itself been deleted, this item is restored only if its head is a verb or a deverbal noun or adjective of a fully productive type. Examples of a coreferential zero subject in the MC (we mark the added nodes by square brackets) are given in (16) and (17): (16) Udělala jsem [já.FEM.SG.ACT.ELID] to. He has done it . (17) Byla [ona.FEM.SG.PAT.ELID] předběhnuta několika jinými. She was left behind by some others . While with (16) the Gender value is based on intrasentential context (the properties of the verb), with (17) a clue is only present in intersentential context: ona (which has been restored on the basis of the grammatical agreement of the verb forms) has also the values of NEUT.PL (e.g. if the neuter noun děvčata 'girls' is the antecedent). Both in MC and in LC, the value of Number with most of the restored subject pronouns will be supplied in the second phase of the automatic procedure, on the input of which the subjectverb agreement has been specified. However, Gender and functor are filled in manually, which is necessary also in case the pronoun has not been deleted. Only in certain specific cases an automatic solution is possible, e.g. with certain occurrences of the Vocative case accompanying the subject, as in (18), or with the verb-subject agreement disclosing the Gender of the subject, as in (19): (18) Ty jsi, Martine, spal? 'You, Martin, have been sleeping?' ty.ANIM.PL.ACT;COREF:Martin;CORNUM:3 jsi, Martine, spal? (19) My jsme tam byly všechny. We (women, girls) have been there all . My.FEM.PL.ACT jsme tam byly.FEM.PL všechny. In examples such as (20), the values of the 'coreferential' attributes should be assigned manually if there is a univocal antecedent in the previous sentence (otherwise just symbols for empty values are present), see also the note following ex. (13') above. It may be recalled that a verb such as prší 'it rains' has no dependent ACT; its valency only admits adverbial adjuncts. (20) Předseda vlády řekl, že předloží návrh na změnu volebního systému. 'The Prime-minister said that (he - the Prime-minister, the Government, or somebody else identifiable on the basis of the context) will submit a proposal on the change of the election system'. These and other similar examples show that the difference between the presence of a Nominative form of a personal pronoun (including the anaphoric 3rd person pronouns) and its ellipsis is not crucial. In either case it is necessary to specify whether a specific antecedent can be found in the context or only an indistinct anaphoric relation can be stated. In the latter case the pronoun may be understood prototypically to refer to the most salient element of the hearer's memory (according the speaker's assumptions), cf. the discussion of ex. (4) in Sect. 2.2 above. Let us remark that textual coreference covers also wider anaphoric relations, which do not represent full referential identity, as e.g. in (21), in which oni 'they' is interpreted as referring to a group that includes George. Up to now, such cases have not been accounted for by the procedure discussed. (21) Jirka zase nepřišel. Oni všichni často chybějí. 'George failed to turn up again. They all often are absent.' 3.4 Specific aspects of deletion Our preliminary analysis of the Czech National Corpus indicates that the following types of deletions have to be recognized, both of which are of an anaphoric character: (i) deletions licensed by the grammatical properties of sentence elements or sentence structure, (ii) deletions possible only if the preceding context (be it co-text or context of situation) exhibits certain specific properties. In both cases the task is not only to insert into the sentence structure a node that has no overt lexical counterpart (not even one that would consist in an occurrence of a function word), but also to specify the antecedent of a 'silent' anaphora. Our subclassification of reconstructions of nodes can be compared with the kinds of 'silent' anaphora in the annotation scheme of the FrameNet project (Fillmore 1999), where especially the counterpart of the case (b) below (that of a "zero morph") has been elaborated in detail. We do not go so far e.g. in the analysis of deverbative nouns, i.e. we just exclude the Actor from the valency frame of an agentive noun, such as writer, instead of characterizing the suffix as filling this slot. In the MC, the arguments (inner participants) of verbs and deverbative nouns are restored, even if their head itself has been deleted and has to be added. In the LC, we in principle do not restore any (deleted) complementations of nouns except for the case of the maximally productive deverbatives with the protypical suffix -á/aní, -tí (e.g. čekání 'waiting' from the verb čekat 'to wait'; we distinguish between psaní 'a letter' and 'writing' as in Dostali jsme psaní 'We got a letter' and Psaní mu trvalo hodinu 'Writing took him an hour'). Several cases of ellipsis (deletion) resolution have been discussed in Sections 3.2 and 3.3 above, since their treatment either is analogical to that of anaphoric items (e.g. with the zero subject in Czech as a pro-drop language it is not crucial whether the subject pronoun has been deleted or not, since the ambiguity of number and gender, and thus the semantic specificity of the anaphoric pronoun are basically the same as those of the verb form exhibiting grammatical agreement with the subject, see the discussion of ex. (17) in Section 2.3 above)2 or is fully based on grammatical conditions given in the sentence (as with control, see Section 3.2.(iii)). In the present section we concentrate on the identification of the positions in which a node is to be inserted into the tree. Within group (i), two situations may obtain: (a) The position itself in the sentence structure is predetermined (i.e. a sentence element is subcategorized for this position), although its lexical setting is 'free'. This is e.g. the case given by the so-called pro-drop character of a language like Czech, where the position of the subject of a verb is 'given', but it may be filled in dependence on the context, cf. the discussion of ex. (11) in Section 3.2 above. We assume that the subject node is to be added in the 'Wackernagel position', since the zero form of the subject pronoun in Czech is used under similar conditions as the weak, clitic pronouns, for which this position is typical. In both cases the relevant item always is contextually bound, i.e. primarily it belongs to the topic of the sentence. Also other nodes that are restored as dependent on a verb get into the 'Wackernagel position' of the given clause. A node restored as dependent on a noun (or adjective) is always placed to the left of its governing word. If more than one node are inserted into one and the same place, then their order should conform to the systemic ('canonical') ordering of the valency slots (functors), i.e., in Czech, ACT is followed by most of the free modifications and then by ADDR, PAT and EFF in this order (see Sgall et al. 1986; 1995). As was illustrated in Section 3.2, a personal pronoun (including the anaphoric 3rd person pronoun) is added as the lexical value of a restored subject node, and its values of Gender and Number are specified according to the congruent form of the verb and to what has been understood from the intra- or intersentential context. Every restored node obtains a functor both in MC and LC, and is marked as deleted (elided) by the value ELID. Also the ellipsis of another semantically obligatory but deletable complementation of a verb is restored in a similar way: e.g. the Cz. verb přijet 'to arrive' has as its obligatory complementation an Actor and a Directional-3, "where-to," (the obligatoriness of the Directional complementation can be tested by the 'dialogue test', see Panevová 1974; Sgall et al. 1986), which can be deleted on the surface, cf. (22); here the item with the functor Directional has been deleted because the speaker assumes that the hearer will identify the referent easily. The deleted item can be restored either as here or there, either of which can be identified in some cases as referring to a specific location. With (22) here may be restored e.g. if the discourse takes place among people waiting for a train at a railway station, while there would be appropriate e.g. after a request according to which the addressee should wait for someone at a station (in both cases the adverbs then would refer to the station). (22) Vlak přijede v šest hodin. 'The train will arrive at six o'clock.' (b) Both the position and its 'filler' is predetermined; this situation might be described as the presence of a 'zero morph' rather than deletability, especially in case the deletion is obligatory. An example of the function of a zero morph are the so-called General Participants, see examples (23) through (25), in which the restored node (or more precisely, the values of its lemma and functor) is included in square brackets. (23) Ta kniha [Gen.ACT.ELID] byla už vydána dvakrát. 'The book has already been published twice.' (24) V neděli [Gen.PAT.ELID] obvykle peču. 'lit.: On Sundays (I) usually bake' (25) Dědeček [Gen.ADDR.ELID] vypravuje pohádky. 'Grandfather tells fairytales' The lemma Gen applies also with adjectives: (26) Jan je věrný [Gen.ADDR.ELID] (manžel). 'John is (a) faithful (husband).' (27) Je to hrdý [Gen.PAT.ELID] člověk. 'He is (a) proud man.' Also the phenomena of control belong here, see Section 3.2 above. Thus, if the types of ellipsis are compared with the two kinds of coreference discussed in Sect. 3.2 and 3.3, it can be seen that whenever ellipsis functions as grammatical coreference, it belongs to group (i)(b). Group (ii) consists of ellipses conditioned by the context, with which both the posItion of the deleted node and its lexical value are determined by the context alone, rather than by the sentence structure. As we have mentioned in our introductory remarks in Sect. 2.1 and illustrated by examples (1) through (3), this is a point in which the typological differences between English and a language with rich inflection, such as Czech, are most clearly to be seen; the rich morphemics allows for deletions in many cases in which a deletion is impossible in the English text. Thus, the deletion of the whole topic of the sentence is possible e.g. in (28): (28) (Potkal jsi včera Toma?) Potkal. 'lit.: (Did you meet Tom yesterday?) Met'. Along with these rather specific instances (in which the verb in a typical context does require the Objective to be present also in the surface), also the following cases are characteristic of contextual deletion in Czech: (a) The restored node is the governor (the head) of a congruent adjective. In this specific case the restored node obtains the position to the left of the adjective, but to the right of its own head word: (29) Přišli jen [ten.ACT.Plur.ELID] mladší. Only (the) younger [ones] came.' (30) Našli jen [ten.PAT.Plur.ELID] modré. They only found (the) blue (ones). A different situation is present with those adjectival words with which we assume a substantival function as well. This is the case of such pronouns as ten(to) 'this', některý 'some', cardinal numerals, superlatives, and of course the 'substantivized adjectives' (nemocný, ill', raněný 'wounded , etc.). As illustrated by (31) and (32), no deletion is present and such adjectival words get their functor themselves, as directly depending on the verb. (31) Zvolili tři.PAT z pěti místopředsedů. They elected three.ACT from the five vice-presidents (32) (Připravili večeři pro deset hostů.) Přišli jen čtyři.ACT. '(They prepared dinner for ten guests.) There came only four.ACT.' Furthermore, a noun is not restored with adjectival words in constructions with the functor PAT with a copula (Kluci byli úspěšní.PAT the boys were successful.PAT ) or with EFF and COMPL: e.g., pokládat za své.EFF regard as (one s) own.EFF , našli je nemocné.COMPL they found them ill.COMPL . (b) Deleted pronouns of laziness may be seen in examples such as (33): (33) Můj bratr políbil svou ženu a já taky. Lit.: My brother kissed his wife and I too. 'My brother kissed his wife and so did I.' Můj bratr políbil svou ženu a já [on.PAT.FEM.SG.ELID políbil] taky. In case it can be found on the basis of context whether 'my own' or 'his wife' is referred to, this is distinguished in the MC by (34) and (35), respectively, where the values in square brackets are those of COREF. (34) já.ANIM.SG.APP bratr.ANIM.SG.ACT políbit.PRET...CO se.ANIM.SG.APP.[bratr] žena.FEM.SG.PAT CONJ já.ANIM.SG.ACT já.ANIM.SG.ELID.APP on.FEM.SG.ELID.PAT.[žena] políbit.PRET...CO taky... (35) já.ANIM.APP bratr... políbit... ...žena.FEM.SG.PAT CONJ on.ANIM.SG.ELID.APP.[bratr] žena.FEM.SG.PAT To make these examples more transparent, we give their version with English lemmas in (34') and (35'): (34') I.ANIM.SG.APP brother.ANIM.SG.ACT kiss.PRET...CO he-Refl.ANIM.SG.APP.[brother] wife.FEM.SG.PAT CONJ I.ANIM.SG.ACT I.ANIM.SG.ELID.APP he.FEM.SG.ELID.PAT.[wife] kiss.PRET...CO also... (35') I.ANIM.APPURT brother... kiss... ...wife.FEM.SG.PAT CONJ he.ANIM.SG.ELID.APP.[brother] wife.FEM.SG.PAT Note: The abbreviations for the individual values are self explaining, except for APP, which stands for the relation of Appurtenance (i.e. 'Possession' in a broad sense), CONJ, which stands for the Conjunction, and CO, which indicates the coordinated items. These specific symbols serve to reduce the more-dimensional structures reflecting the combinations of dependency and coordination to bidimensional tree structures. A pronoun of laziness that has not been left out remains as a pronoun. For the time being, the value of its COREF is filled in only in the MC. (c) In coordination structures nodes for the deleted repetitions of the governing word are restored under the condition that the resulting structure will be as simple as possible, if this is not excluded by clear semantic or syntactic factors. Thus, (36) is handled as not including a deletion, since the two coordinated predicates can be understood as to be 'jointly' modified by the two arguments. (36) Jirka potkal a pozdravil Marii. George met and greeted Mary Sometimes adding a node is inevitable: in (37) the presence of deletion is clearly given by the fact that some of the dependents of the two heads (one of which is deleted) differ: (37) Potkal Marii včera a já [jsem Marii potkal] dnes. He met Mary yesterday and I [met Mary] today. In such a case, the verb is restored in the manual phase, but its Objective is then specified by the second part of the automatic procedure, which identifies it in accordance with the lefthand branch of the coordination construction. The lemma of the new node gets the shape of a noun, rather than a pronoun, since coreferentiality is present in all readings of such a sentence. A different situation obtains with e.g. (38), in which the pronoun ji 'her' can refer to another person, if this person has been strongly activated by the context: (38) Potkal Marii včera a já jsem ji potkal dnes. 'He met Mary yesterday and I met her today.' In other cases the inevitability of restoring a deleted item is given semantically, cf. (39), in which the two adjuncts cannot be interpreted as depending on a single occurrence of the noun: (39) Pil červené a bílé víno. 'He drank red and white wine'. We are aware of the difficulties connected with drawing such a boundary between sentences with and without deletion. On the one hand, the theory of language cannot exclude the possibility that once a kind (or way of existence) of wine comes about that would somehow adopt two colors, cf. e.g. such noun groups as a red and white flag. On the other hand, there are cases with which the annotators have to look for clues in a broader context, and perhaps do not find them (cf. the much discussed example of old men and women). However, as far as practical issues of natural language processing are concerned, the present preliminary solution seems to be relatively suitable, at least before very large sets of examples can be studied (which will only be possible on the basis of very large syntactically tagged corpora). It does not seem to be crucial that some of the cases under (b) also meet the conditions of (a) above. If along with the node that is being restored, also one or more nodes depending on it are deleted, then both in the LC and in the MC these are restored only if they function as arguments (rather than adjuncts) of their head; adjuncts ('free complementations') are restored only if they are obligatory with the given head. In MC the symbol ELEX is distinguished from ELID: ELEX is assigned to the restored node when its optional adjuncts are not restored in spite of the fact that they should be understood as deleted (i.e. present in the meaning of the sentence). Thus, e.g., with (40) the verb is restored, but the optional adjunct včera 'yesterday' is not restored in the second part of the conjunction; the existence of a further modification that could be transferred from the first clause gets merely indicated by adding navštívit.CO.ELEX as the rightmost member of the coordination structure). (40) Včera navštívil Jirka Marii a Milan Annu. Lit: 'Yesterday George visited Mary and Milan Ann.' 4. Conclusions and prospects Neither the automatic nor the manual part of the tagging can achieve a complete formulation of tectogrammatical representations. Although high precision can be achieved over unrestricted texts - in several respects (e.g. in what concerns anaphoric pronouns within a single sentence, especially reflexive pronouns, or ellipsis with control and coordination), other types of coreference can be resolved only after further empirical investigations. Our brief characterization of how the deleted nodes are restored in PDT could only illustrate some aspects of the first steps of the procedure of semi-automatic syntactic tagging. For most further steps it will be possible to use (perhaps also with applications of statistical methods) the large set of annotated utterances obtained in this way. The existence of such a set will make it possible to check the solutions proposed here on the basis of monographic studies, which will find a much more suitable starting point with the syntactically tagged corpus than with the previously used excerpts. With a large corpus tagged not only on the level of morphemics, but also on that of (underlying) syntax, theoretical linguistics will be enabled to check its descriptive framework; this is quite a new situation for our branch of science. We have seen that it is impossible to handle ellipsis as entirely distinct from anaphora. In both cases, relevant grammatical factors are to be distinguished from contextual ones, and the resolution can only be looked for in a complex procedure. While certain parts of this procedure can already be included into automatic modules of syntactic tagging, other steps still require intellectual analysis. Only when a set of utterances will be large enough to serve as a training set for automatic procedures containing both structural and statistical ingredients, a relatively full automation of the whole procedure can be achieved. The approach characterized in the present paper, enriched by a procedure that makes it possible to trace the development of the degrees of salience (see Hajičová et al. 1981; 1995), will make it possible to specify not only the coreferential links within a sentence, but also intersentential relationships. Such an approach to the analysis and description of natural language seems to offer a sound basis for a procedure going the whole way from the outer form of texts to an appropriate form of the input for semantico-pragmatic interpretation. Acknowledgement: The research reported on in this paper has been predominantly carried out within the project supported by the Czech Grant Agency 405-96-K214, and in part by that of the Czech Ministry of Education VS 96-151. Notes: 1 Although there are anaphoric expressions that are not strictly coreferential (being connected with associative relations such as whole-part or set-element), we use the term 'coreference' as covering both the use of anaphoric means (pronouns, nouns and other items) and ellipsis. With the latter, the deleted item always has an anaphoric function, as will be documented by the discusion that follows. 2 Only in certain specific cases (restricted to a subset of gender or number values, or of verb classes) the verb forms display a higher ambiguity, e.g. in PL the finite verb forms do not distinguish Gender; forms such as trpí 'suffer(s)' agree both with on 'he/she/it' (SG) and with oni 'they' (PL). References: Böhmová A. and E. Hajičová (1999). The Prague Dependency Tree Bank I: How much of the underlying syntactic structure can be tagged automatically? The Prague Bulletin of Mathematical Linguistics 71, 5-12. Fillmore C. J. (1999). Silent anaphora. Paper presented at the TELRI Workshop, November 1999, Bratislava. Hajič J. (1998). Building a syntactically annotated corpus: The Prague Dependency Treebank. In: Issues of Valency and Meaning. Studies in Honour of Jarmila Panevová, ed. by E. Hajičová, 106-132. Prague: Karolinum. Hajič J. and B. Hladká (1997). Morfologické značkování korpusu českých textů stochastickou metodou [Morphological tagging of Czech using stochastic methods] Slovo a slovesnost 58:288-304. Hajičová E. (1993). Issues of sentence structure and discourse patterns. Prague: Charles University. Hajičová E. (1998). Prague Dependency Treebank: From analytic to tectogrammatical annotations. In: Text, Speech, Dialogue (eds. Petr Sojka, Václav Matoušek, Karel Pala and Ivan Kopeček), Brno: Masarykova univerzita, 45-50. Hajičová E., Hoskovec T. and P. Sgall (1995). Discourse modelling based on hierarchy of salience. Prague Bulletin of Mathematical Linguistics 64, 5-24. Hajičová E., J. Panevová and P. Sgall (1999). Manuál pro tektogramatické značkování Českého národního korpusu [Manual for tectogrammatical tagging of the Czech National Corpus], Tech. Report No. 7, Institute of Formal and Applied Linguistics, Charles University, Prague. English translation to be published in 2000. Hajičová E. a J. Vrbová (1981), On the salience of the elements of the stock of shared knowledge. Folia linguistica 15:291-303. Hladká B. (2000). Czech language tagging. Ph.D. dissertation, Charles University, Prague. Panevová J. (1974). On verbal frames in Functional Generative Description. Prague Bulletin of Mathematical Linguistics 22:3-40, 23(1975):17-52; a revised version in Prague Studies in Mathematical Linguistics 6, 1978, 227-254. Panevová J. (1991). Koreference gramatická nebo textová? [Grammatical or textual coreference?] In: Etudes de lingustique romane et slave, ed. by W. Banys, L. Bednarczuk, and K. Bogacki), Cracow, 495-506. Petkevič V. (1995). A new formal specification of underlying structures. Theoretical Linguistics 21:7-61. Sgall P. (1992). Underlying structure of sentences and its relations to semantics. Wiener Slawistischer Almanach. Sonderband 33. Ed. by T. Reuther, 273-282. Sgall P., E. Hajičová and J. Panevová (1986). The meaning of the sentence in its semantic and pragmatic aspects, ed. by J. L. Mey. Dordrecht:Reidel - Prague:Academia. Sgall P., O. Pfeiffer, W. U. Dressler, and M. Půček (1995): Experimental research on Systemic Ordering. Theoretical Linguistics 21:197-239.